Automated Classification of Kidney Diseases Using Deep Learning: A Comparative Analysis with Machine Learning Models on CT Images

Kidney disease diagnosis with the use of a medical image plays an important role in diagnosis and treatment on time. This survey presents a comparison between different machine learning algorithms and deep learning models that classify kidney conditions from the CT KIDNEY DATASET, containing four distinct classes: Normal, Cyst, Tumor, and Stone. The authors have reviewed the performance of CNNs while comparing them against traditional machine learning models like Decision Trees, Random Forests, Gradient Boosting, AdaBoost, Support Vector Classifiers (SVC), Logistic Regression, K-Neighbors, Naive Bayes, and Long Short-Term Memory (LSTM) networks. The best performance, by the CNN model, is 93.87%, which states the power of deep learning in image processing regarding medical applications. In contrast, Random Forest and Decision Tree models remain competitive within traditional classifiers with validation accuracies of 83.37% and 83.05%, respectively. Gradient Boosting, AdaBoost, and SVC models come next with validation accuracies ranging from 82.73% to 79.60%, whereas KNeighbors and Naive Bayes are still farther behind, although each was performing consistently. It reflects the particular challenges associated with the application of recurrent models to static image data, best suited for temporal data. Our LSTM model realizes a test accuracy of $\mathbf{75.18 \%}$. It also points to the fact that CNNs exhibit a far better capability in detecting kidney diseases, whereas traditional models are quite viable alternatives with competitive accuracy. In fact, this essentially points out that model selection is an important issue in clinical applications, as accuracy and computational efficiency go hand in hand to enable accurate diagnosis.